Spaces:
Paused
Paused
| title: Text Embedding Explorer for SD 1.5 | |
| colorFrom: blue | |
| colorTo: indigo | |
| sdk: gradio | |
| sdk_version: 6.30.0 | |
| app_file: app.py | |
| suggested_hardware: t4-small | |
| short_description: Explore and edit text embeddings in Stable Diffusion 1.5 | |
| pinned: false | |
| # Text Embedding Explorer for Stable Diffusion 1.5 | |
| A Gradio Space for BA creative coding students. Instead of changing the prompt text, you change | |
| the **text embedding** (the 77 x 768 numbers that Stable Diffusion 1.5 actually receives). | |
| Tabs: | |
| 1. **Interpolate** blends two prompts (lerp or slerp, with optional extrapolation). | |
| 2. **Arithmetic** computes A - B + C on embeddings. | |
| 3. **Edit tokens** scales or blends single tokens of a prompt and shows original and edit side by side. | |
| 4. **Map** draws a 2D PCA map and a similarity matrix of words or prompts. | |
| All images within one run share the same noise (seed), so differences come from the embedding only. | |
| ## Hardware | |
| Made for a dedicated GPU such as `t4-small`. Typical time on a T4 is a few seconds per image at 512 px | |
| and 20 steps. On CPU it runs but takes minutes per image. | |
| ## Settings | |
| Optional environment variables (Space settings, Variables): | |
| - `MODEL_ID` loads another SD 1.5 compatible repo. Default is `stable-diffusion-v1-5/stable-diffusion-v1-5`. | |
| - `DISABLE_SAFETY_CHECKER=1` turns off the NSFW filter. It is on by default. Odd embedding mixes can | |
| occasionally trigger it, which shows up as a black image. | |
| ## Run locally | |
| ```bash | |
| pip install -r requirements.txt gradio==6.30.0 | |
| python app.py | |
| ``` | |
| The model is downloaded from the Hugging Face Hub on first start. | |